Add support for allocating memory that has the same iova as its virtual address#58
Open
lsgunth wants to merge 13 commits into
Open
Add support for allocating memory that has the same iova as its virtual address#58lsgunth wants to merge 13 commits into
lsgunth wants to merge 13 commits into
Conversation
.dma_unmap() takes an iova followed by the length, but the length was erroneously specified first. This was caught while creating the next two patches. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
birkelund
reviewed
May 22, 2026
|
|
||
| #define __static_assert(x) static_assert(x, #x) | ||
|
|
||
| typedef uint64_t iova_t; |
Collaborator
There was a problem hiding this comment.
Can this be moved to include/vfn/iommu/dma.h without too much trouble?
Author
There was a problem hiding this comment.
Thanks!
I'll have to double check next week, but I believe I put it in support because it was a little awkward to have it defined in iommu/dma.h when it was needed by most of the headers in iommu which didn't need to include eachother.
I'll try changing it next week and if it's not too problematic I'll update this PR.
Instead of using a plain uint64_t to refer to iovas, use a specific type (iova_t) to make the code a bit more self documenting. This can make some useages a bit clearer. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Adding the nocast attribute to iova_t lightly prevents the type from being automatically casted to other types. This helps ensure that iova_t is used in the correct places at the cost of needing an occasional cast if doing math with the iova or passing an iova to the skiplist. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Store a convient way to get the maximum iova value. This will be useful to create buffers with the same iova and virtual address. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add helper function to allocate memory where the iova and the virtual address are the same. While the skiplist to translate a virtual address to an iova is relatively performant, in some applications it introduces extra undesirable overhead. To avoid this overhead, it is desirable to be able to allocate memory where the iova and the virtual address are the same. To do this, mmap() memory with a hint to locate it at a region starting at a quarter of the maximum iova value. Then use IOMMU_MAP_FIXED_IOVA to assign an iova that matches the virtual address. If the mmap() fails (which is pretty unlikely) to assign an address in the desired region, increment the next iova to attempt and retry. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
In all cases, the callers of nvme_mapv_prp() and nvme_mapv_sgl() know the iova address of the segment as they always typically come from the request. Instead of looking these up in the skiplist, just pass the known value directly. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Introducing nvme_mapv_iova_prp() and nvme_mapv_iova_sgl() which take an iova_vec pointer that can be used to avoid calling iommu_translate_vaddr() when mapping vectored commands. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add simple wrappers for nvme_rq_mapv_iova_prp() and nvme_rq_mapv_iova_sgl() similar to the non-iova versions. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add the -S flag to the io example that, when specified, allocates the memory buffer using memory that has the iova as virtual address. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Replace several 0x1000 values with a single io_sz variable. This will allow changing the size in the next patch. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add the -v flag for vectored SGL and the -s flag for sgl. This help testing the nvme_rq_mapv_[prp|sgl]() functions. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add the -i/--iova option to test passing iova values directly to the vector functions to avoid the skip list. Signed-off-by: Logan Gunthorpe <logan.gunthorpe@eideticom.com>
Add tests for nvme_rq_mapv_iova_prp() and nvme_rq_mapv_iova_sgl(). These are duplicates of the non-iova versions.
Author
|
I was able to move those defines without any difficulties. I've force pushed with it updated. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
In some situations it may not be practical to pass around an iova address along with its virtual address. To support these situations this PR aims to add a way to allocate memory where the iova is the same as the virtual address. It does this by dedicating a region in the iova address space well above the addresses that would be used by existing code and using the MAP_FIXED_NOREPLACE flag with mmap() to attempt to allocate memory with an address in this region.
In addition to the core feature, this PR also adds type definitions and checking with sparse around iova addresses, fixes a parameter swap bug that was caught by the extra type checking, and adds some improvements to the examples to help demonstrate and test the feature.